Paper Diagnoses 'Forecast Collapse' in TSFM Stock Ranking

Point a time-series foundation model at hourly returns for 1,000 US equities and its predictions "become nearly flat and show poor stock ranking, as measured by cross-sectional correlation." That failure mode is what a new paper on arxiv calls "forecast collapse."

The same models handle trading volume in the same setting without much drama. That contrast is what lets the authors, Shu Wan, Miles Ma, Hank Zhu, Guangqi Liu, Stephen Wang, Qingsong Wen and Huan Liu, argue the failure is "closely tied to target predictability" rather than a bug in any single architecture. Their investigation spans TSFMs, twelve deep-learning forecasting models, and 97 public benchmark configurations.

The paper attributes the collapse to two things at once: low predictability limits the amplitude of calibrated point forecasts, and per-series training objectives "leave cross-series structure unidentified." Push in either direction alone and something breaks. Optimising squared error flattens the predictions; directly optimising cross-sectional correlation restores ranking but can inflate forecast amplitude "by more than an order of magnitude."

The authors' proposed fix, CalibRank, is an objective that trades calibration against ranking. On their Finance1K dataset, they report it "nearly triples cross-sectional correlation while keeping amplitude close to the target," and improves correlation on every model tested.

The wider claim is a critique of how the field measures itself: "per-series metrics can hide failures in cross-series structure needed by downstream decisions."